DeerFlow – an open‑source super‑agent platform

What it is – DeerFlow (Deep Exploration and Efficient Research Flow) is a framework for building AI‑agents that can call other agents, use tools, keep memory, and run in isolated sandboxes. It is positioned as a “super‑agent harness” that lets developers assemble sub‑agents, skills, RAG (private knowledge retrieval), and browser‑based web‑search into a single, configurable workflow.

---

Core ideas & why it matters

Idea	What it gives you
Super‑agent harness	A central “lead‑agent” can orchestrate many sub‑agents (e.g., a planner, a coder, a data‑retriever) and pass results between them.
Extensible skills / tools	Built‑in tool adapters for web search (InfoQuest, Jina, Browserless, etc.), file‑system sandbox, code execution, and custom Python‑based skills. New skills can be added by exporting a function.
Memory layers	Short‑term “session goals” and “current‑task notes”, plus a long‑term vector store called RAGFlow for private knowledge retrieval.
Safety sandbox	Optional sandbox mode isolates the agent’s OS access, preventing arbitrary code execution unless explicitly enabled.
Tracing & observability	Native integrations with LangSmith, Langfuse, Monocle, and request‑trace correlation for debugging complex multi‑step runs.
Multi‑model support	Configure any OpenAI‑compatible model (OpenAI, Gemini, DeepSeek, Claude, Qwen, vLLM, etc.) and even mix providers in a single run.

These pieces let you build sophisticated autonomous workflows—think research assistants that browse the web, retrieve internal documents, write code, and keep a persistent knowledge base—without wiring everything by hand.

---

Getting started (quick‑start summary)

1. Clone & run the wizard

   git clone https://github.com/bytedance/deer-flow.git && cd deer-flow
   make setup   # interactive wizard creates config.yaml & .env
   

2. Choose a model – the wizard lets you add OpenAI, Gemini, DeepSeek, Claude, or a local vLLM endpoint.
3. Pick optional web‑search – enable InfoQuest, Jina, Browserless, etc.
4. Start the system – either make dev for local dev or the recommended Docker compose (make docker).
5. Interact – open the web UI (or LangGraph Studio) and begin a chat; the lead‑agent will automatically invoke sub‑agents, tools, or RAG as needed.

The repo also ships a Terminal Workbench (TUI) for headless use and a Python client to embed DeerFlow in other code.

---

Main features (as listed in the README)

• Skills & Tools – export custom Python functions as “skills”; built‑in Claude‑Code integration.
• Private Knowledge Retrieval (RAGFlow) – vector store for your own documents.
• Chat Archive – persistent conversation history.
• Session Goals & Notes – structured context that survives across turns.
• Sub‑Agents – hierarchical agents that can be limited per‑run.
• Sandbox & File System – optional isolated execution environment.
• Agentic Browser Control – browse, click, and scrape via the InfoQuest crawler.
• Context Engineering – manual compaction, prompt overlays, and request‑admission throttling.
• Tracing – LangSmith, Langfuse, Monocle, and custom request‑trace IDs.
• Multi‑Provider Models – configure dozens of models in config.yaml or via the UI.
• Scheduled Tasks – cron‑like jobs that can trigger agents automatically.
• Terminal Workbench – a TUI for managing agents without a browser.

---

Who should use it?

• AI product teams that need a plug‑and‑play platform for building autonomous assistants.
• Researchers exploring multi‑agent coordination, tool‑use, or RAG pipelines.
• Developers who want a single place to configure models, sandbox execution, and observability.
• Enterprises looking for an open‑source alternative to proprietary agent frameworks.

---

Limitations & cautions (from the security notice)

• Deploying without the recommended sandbox can give the agent unrestricted shell / file‑system access.
• The “gateway admin” role effectively grants code‑execution rights.
• Mis‑configured web‑fetch back‑ends may expose internal networks; the system blocks private‑address fetches by default.
• Cost‑estimation only works when all models use the same currency.

---

License & community

• MIT License – permissive for commercial and academic use.
• Active community (the repo hit #1 on GitHub Trending in Feb 2026). Contributions are welcomed on the 2.0 branch; the older 1.x branch still hosts the original Deep Research framework.

---

TL;DR

DeerFlow is a full‑stack, open‑source platform for building autonomous, tool‑using AI agents with memory, web‑search, and multi‑model support. It provides a Docker‑ready deployment, a web UI, a TUI, and extensive tracing, making it a practical foundation for research or product‑grade AI assistants.